Eval: edit-quality cases for b2_edit (clean match, preservation, judged quality, honesty) #257

Part of #255 (ADR-0023, ADR-0013). Read docs/evals.md process rules before touching corpus, labels or metrics.

Cases

• Clean match rate. Over a set of "clean up / tighten / fix headings" prompts on corpus notes, the share of b2_edit calls that apply without NoMatch or Ambiguous.
• Preservation. After applying all proposals, every [[wikilink]], heading and frontmatter key the note had is still present unless the prompt asked to remove it.
• Judged quality. A judge scores the resulting note against the prompt, as answers are scored today.
• Honesty. A prompt with nothing to fix yields zero proposals and an answer that says so.
• Focus. The Chat doesn't know which note is open: "clean up this doc" gets "I don't find that in your notes" #253 case: "this doc" with focus answers from the focused note; without focus it still gets the no-evidence sentence.

Decide

Report per model tier (PR #250). The result decides whether edit mode is on by default for the shipped model or gated by tier. Record the run in docs/evals.md.